aiokdb 0.1.33

Pure Python asyncio connector to KDB

aiokdb

Python asyncio connector to KDB. Pure python, so does not depend on the k.h bindings or kdb shared objects, or numpy/pandas. Fully type hinted to comply with PEP-561. No non-core dependencies, and tested on Python 3.8 - 3.12.

Peer review & motivation

qPython is a widely used library for this task that maps KDB tables to Pandas Dataframes. Sometimes a dependency on numpy/pandas is not desired, or the dynamic type mapping might be unwanted.

This library takes a different approach and aims to replicate using the KDB C-library functions, ie. being 100% explicit about KDB types. It was built working from the publicly documented Serialisation Examples and C API for kdb+ pages. Users might also need to be familiar with k.h.

% python
Python 3.9.22 (main, Apr  8 2025, 15:21:55)
>>> from aiokdb import *
>>> from aiokdb.extras import ktns, ktni, ktnb
>>> x = xt(xd(ktns("hi", "there"), kk(ktni(TypeEnum.KI, 45, 56), ktnb(False, True))))
>>> x  # shows repr() output
xt(xd(ktns('hi', 'there'), kk(ktni(TypeEnum.KI, 45, 56), ktnb(False, True))))
>>> from aiokdb.format import AsciiFormatter
>>> print(AsciiFormatter().format(x))
hi there
--------
45 0
56 1
>>> len(x)
2
>>> x[0]
xd(ktns('hi', 'there'), kk(ki(45), kb(False)))
>>> print(AsciiFormatter().format(x[1]))
hi   | 56i
there| 1
>>> x[2]
IndexError

Basic RPC, using blocking sockets:

# run ./q -p 12345 &

from aiokdb.socket import khpu

h = khpu("localhost", 12345, "kdb:pass")

# if remote returns Exception, it is raised here, unless khpu(..., raise_krr=False)
result = h.k("2.0+3.0")

assert result.aF() == 5.0

result.aJ() # raises ValueError: wrong type KF (-9) for aJ

The result object is a K-like Python object (a KObj), having the usual signed integer type available as result.type. Accessors for the primitive types are prefixed with an a and check at runtime that the accessor is appropriate for the stored type (.aI(), .aJ(), .aH(), .aF() etc.). Atoms store their value to a bytes object irrespective of the type, and encode/decode on demand. Atomic values can be set with (.i(3), .j(12), .ss("hello")).

Arrays are implemented with subtypes that use Python's native arrays module for efficient array types. The MutableSequence arrays are returned using the usual array accessor functions .kI(), .kB(), .kS() etc.

 kdb type name       python    python    python       python  python
  n   c              TypeEnum  accessor  setter       create  type
 ------------------------------------------------------------------------------------
 -19  t   time       -KT       -         -            -       -
 -18  v   second     -KV       -         -            -       -
 -17  u   minute     -KU       -         -            -       -
 -16  n   timespan   -KN       -         -            -       -
 -15  z   datetime   -KZ       -         -            -       -
 -14  d   date       -KD       -         -            -       -
 -13  m   month      -KM       -         -            -       -
 -12  p   timestamp  -KP       -         -            -       -
 -11  s   symbol     -KS       .aS()     .ss("sym")   ks()    str
 -10  c   char       -KC       .aC()     .c("c")      kc()    str (len 1)
  -9  f   float      -KF       .aF()     .f(6.1)      kf()    float
  -8  e   real       -KE       .aE()     .f(6.2)      ke()    float
  -7  j   long       -KJ       .aJ()     .j(7)        kj()    int
  -6  i   int        -KI       .aI()     .i(6)        ki()    int
  -5  h   short      -KH       .aH()     .h(5)        kh()    int
  -4  x   byte       -KG       .aG()     .g(4)        kg()    int
  -2  g   guid       -UU       .aU()     .uu(UUID())  kuu()   uuid.UUID
  -1  b   boolean    -KB       .aB()     .b(True)     kb()    bool
   0  *   list        K        .kK()     -            kk()    MutableSequence[KObj]
   1  b   boolean     KB       .kB()     -            ktnb()  MutableSequence[bool]
   2  g   guid        UU       .kU()     -            ktnu()  MutableSequence[uuid.UUID]
   4  x   byte        KG       .kG()     -            ktni()  MutableSequence[int]
   5  h   short       KH       .kH()     -            ktni()  MutableSequence[int]
   6  i   int         KI       .kI()     -            ktni()  MutableSequence[int]
   7  j   long        KJ       .kJ()     -            ktni()  MutableSequence[int]
   8  e   real        KE       .kE()     -            ktnf()  MutableSequence[float]
   9  f   float       KF       .kF()     -            ktnf()  MutableSequence[float]
  10  c   char        KC       .kC()     -            cv()    array.array
  11  s   symbol      KS       .kS()     -            ktns()  MutableSequence[str]
  12  p   timestamp   KP       -         -            -       -
  13  m   month       KM       -         -            -       -
  14  d   date        KD       -         -            -       -
  15  z   datetime    KZ       -         -            -       -
  16  n   timespan    KN       -         -            -       -
  17  u   minute      KU       -         -            -       -
  18  v   second      KV       -         -            -       -
  19  t   time        KT       -         -            -       -
  98      flip        XT       .kkey(), .kvalue()     xt()    KObj, KObj
  99      dict        XD       .kkey(), .kvalue()     xd()    KObj, KObj
 100      function    FN       -         -            KFnAtom       -
 101  ::  nil        NIL       -         -            kNil    -
 127      `s#dict     SD       .kkey(), .kvalue()     -       KObj, KObj
-128  '   err        KRR       .aS()     .ss()        krr()   str

Serialisation is handled by the b9 function, which encodes a KObj to a python bytes, and the d9 function which takes a bytes and returns a KObj.

Calling repr() on KObj returns a string representation that, when passed to eval(), will exactly recreate the KObj. This may be an expensive operation for deeply nested or large tables.

• Atoms are created by ka, kb, ku, kg, kh, ki, kj, ke, kf, kc, ks, kt, kd, kz, ktj
• Vectors from python primitives with ktnu, ktni, ktnb, ktnf, ktns, passing desired TypeEnum value as the first argument.
• Mixed-type objects lists with kk.
• Dictionaries with xd and tables with xt.

Python manages garbage collection, so none of the reference counting primitives exist, i.e. k.r and functions r1, r0 and m9, setm.

Asyncio

Both kdb client and server protocols are implemented using asyncio, and can be tested back-to-back. For instance running python -m aiokdb.server and then python -m aiokdb.client will connect together using KDB IPC. However since there is no interpreter (and the default server does not handle any commands) the server will return an nyi error to all queries. To implement a partial protocol for your own application, subclass aiokdb.server.ServerContext and implement on_sync_request(), on_async_message(), and perhaps check_login().

Command Line Interface

Usable command line client support for connecting to a remote KDB instance (using python asyncio, and prompt_toolkit for line editing and history) is built into the package:

$ pip install aiokdb prompt_toolkit
$ ./q -p 12345 &
$ python -m aiokdb.cli --host localhost --port 12345
(eval) > ([s:7 6 0Nj]x:3?0Ng;y:2)
s| x                                    y
-|---------------------------------------
7| 409031f3-b19c-6770-ee84-6e9369c98697 2
6| 52cb20d9-f12c-9963-2829-3c64d8d8cb14 2
 | cddeceef-9ee9-3847-9172-3e3d7ab39b26 2
(eval) > 4 5 6!(`abc`def;til 300;(3 4!`a`b))
4| abc def
5| 0 1 2 ... 297 298 299
6| KDict
(eval) > [ctrl-D]
$

Formatting

We implement repr(KObj) as a recursively descending, type explicit formatter, suitable for logging complex or unknown responses, albeit whose size could be very large for large datasets. The implementation of str(KObj) is non-descending, and will always be constant time/space irrespective of the payload.

Text formatting (as shown above) is controlled by aiokdb.format.ASCIIFormatter, which looks inside a KObj to render XD, SD, XT types in tabular form containing atom and vector values. Nested complex types ie. dictionary or tables render as the literals KDict or KFlip, so the outer form is preserved. The output can be bounded by passing arguments to the formatter, e.g. ASCIIFormatter(width=120, height=20).

Finally aiokdb.format.HtmlFormatter is suitable for dashboards etc and can be trivially added to web frameworks as a 'template tag' or similar.

QDB Files

Ordinary .qdb files written with set can be read by kfromfile or written by ktofile:

>>> from aiokdb.files import kfromfile, ktofile
>>> k = kfromfile('test_qdb0/test.qdb')
>>> k
<aiokdb.KObjArray object at 0x7559136d8230>
>>> from aiokdb.format import AsciiFormatter
>>> fmt = AsciiFormatter()
>>> print(fmt.format(k))
[5, hello]

Ordinarily k is dictionary representing a KDB namespace containing other objects.

There is no support for splayed or partitioned datasets, however the primitives are (or, at least used to be) the same so this would be possible.

Tests

The library has extensive test coverage, however de-serialisation of certain (obscure) KObj may not be fully supported yet. PR's welcome. All tests are pure python except for those in test/test_rpc.py, which will use a real KDB server to test against if you set the KDB_PYTEST_SERVICE environment variable (to a URL of the form kdb://user:password@hostname:port), otherwise that test is skipped.

• Formatting with ruff check .
• Formatting with ruff format .
• Check type annotations with mypy --strict .
• Run pytest . in the root directory

License